Camera Ballistics from Sensor Imperfections

Grooves made in a gun barrel impart a spin to the projectile for increased accuracy and range. These grooves introduce distinct markings to the bullet fired, and can therefore be used to link a bullet to a specific handgun. In the same spirit, imperfections in manufacturing lead to slight deviations between the precise amount of light that strikes a camera sensor and the recorded pixel values. These deviations vary from camera to camera (even of the same make and model) and can therefore be used for camera ballistics — linking an image to a specific digital camera. 

A digital image is corrupted by sensor imperfections in two basic ways: (1) each pixel value is modulated by a small amount that is independent of the pixel value (this is termed additive noise); and (2) each pixel value is modulated by a small amount that depends on the pixel value (this is termed multiplicative noise, or more formally, the photo response non-uniformity, PRNU). It is this later imperfection that is distinct to a sensor and is used for camera ballistics.

Given either the camera in question, or several images from the camera in question (usually ten or more), the camera-specific PRNU can be estimated using fairly standard image processing techniques. The estimated PRNU can then be compared against a single image to determine if it is likely that it was recorded by the camera in question. This technique has proven to be highly effective and reliable and, amazingly, works even for heavily JPEG compressed images (this technology was used to verify critical evidence in a disturbing child abuse and pornography case in Scotland in 2009).

This basic ballistic technique was first proposed in 1999 [1] and expanded upon and refined in a series of influential papers starting in 2006 [2-4].  If you are interested in digging into the technical details behind this ballistic technique, check out the references below.

1. Kurosawa, Kuroki, and Saitoh, CCD Fingerprint Method-Identification of a Video Camera from Videotaped Images, in IEEE International Conference on Image Processing, pages 537-540, 1999
2. Lukas, Fridrich and Goljan, Digital Camera Identification from Sensor Noise (2006), in: IEEE Transactions on Information Forensics and Security, 1:2(205-214) 
3. Chen, Fridrich, Goljan and Lukas, Determining Image Origin and Integrity Using Sensor Noise (2008), in: IEEE Transactions on Information Forensics and Security, 3:1(74-90)
4. Fridrich, Digital Image Forensic Using Sensor Noise (2009), in: IEEE Signal Processing Magazine, 26:2(26-37)

Comments

[CMC]

Couldn't this be defeated by adding/subtracting equivalent random noise patterns from the pic?

[Because the noise being analyzed is multiplicative in nature, adding/subtracting a random noise pattern will not have much of an impact. It has been argued that an attacker can estimate and remove the sensor pattern noise and insert another pattern. However, the authors of the original forensic technique have developed a counter-counter measure [1]. Stay tuned for a counter-counter-counter measure. -Hany]

[1] Goljan, Fridrich, Chen, "Defending Against Fingerprint-Copy Attack in Sensor-Based Camera Identification", IEEE Transactions on Information Forensics and Security, 6(1):227-236, 2011.